LumeAPILast verified: 2026-08-19. DeepSeek Harness is a developer preview; plugin interfaces may change.
DeepSeek Harness (DSH) is an open-source coding Agent released by DeepSeek on August 13, 2026. It hit 150,000+ GitHub stars in five days, making it one of the fastest-growing developer tools of the year. The core idea is simple: instead of using one model for every task, DSH lets you route different kinds of work to different models---cheap ones for simple edits, premium ones for hard reasoning---all inside a single open-source framework.
This page is for developers evaluating DSH. It covers what DSH can do for you, when it makes sense to add a second model alongside DeepSeek, and how to configure that in five minutes.
What DSH Can Do For You
DSH is not a chat box. It is a coding Agent that runs inside a project directory, reads your files, runs shell commands, calls tools, and keeps a full log of every step so you can rewind and replay sessions.
The four run modes and what they unlock
DSH ships with four run modes. Each one decides which tools the model can use and how much autonomy it has. The right mode depends on the job.
| Mode | What the model can do | Real example |
|---|---|---|
| Standard | Read/write files, run shell commands, search code, plan, spawn subagents, follow skills | "Add JWT authentication to this Express app." DSH reads the routes, adds middleware, updates the login endpoint, writes tests. |
| PTC | Write a TypeScript program that batches multiple tool calls into one execution block | "Refactor every database query in this repo to use prepared statements." The model finds all query files, rewrites them, and runs the test suite in one batch. |
| Minimal | Only shell + file edit---nothing else | "Fix the typo in README.md." One file, one edit, no overhead. |
| Creator | Standard tools plus runtime introspection | "Build a custom preset that always runs ESLint before committing." The model inspects DSH internals and writes a new preset saved to ~/.dsh/.agent-presets/. |
What you get out of the box
- One framework, four autonomy levels. You do not need separate tools for "chat," "code review," and "refactor." DSH covers all three with a single binary.
- Full session replay. Every system prompt, reasoning step, tool call, and result is logged. You can rewind, branch, and replay.
- No vendor lock-in by default. Models are plugins. The default works with DeepSeek, but you can add Claude, GPT, or any OpenAI-compatible model.
- Open source (MIT), local-first. DSH runs on your machine. Your code and your model calls do not leave your environment unless you explicitly send them somewhere.
The gap most teams hit in week two
The default DeepSeek V4 Pro is good enough for the first week. After that, real coding work exposes its limits:
- Code review needs lower hallucination. DeepSeek V4 Pro is decent, but Claude Sonnet has a better track record on careful review work. Many teams keep DeepSeek for generation and Claude for review.
- Large refactors need 1M+ token windows. Refactoring a 200K-token file with DeepSeek V4 Pro is painful. GPT-5.5 and Gemini 2.5 Pro have 1M+ token windows.
- Simple edits should not cost premium-model rates. Autocomplete and one-line changes are 50-100x cheaper on a Flash-tier model.
- PTC mode is stronger on some models than others. Batch migrations work best on models that are good at structured code generation; the best choice depends on the language and codebase.
This is the wall most teams hit. The fix is to add a second model so DSH can route the right work to the right model.
When It Makes Sense to Add Another Model
You do not need a second model on day one. Add one when one of these becomes true for your team:
- You have a monthly bill that hurts. A typical team running DSH Standard mode 4-6 hours a day spends 40-60% of their API budget on tasks a Flash-tier model could handle.
- One model keeps getting things wrong. If your team keeps redoing Claude's code review or fighting DeepSeek's 128K context limit, the answer is not "try harder," it is "use the right model for that job."
- A provider had an outage and your team stopped working. A second provider is your fallback. Single-vendor setups are a single point of failure.
- You want Claude or GPT for a specific task. DSH defaults do not include Anthropic or OpenAI. If your team standardizes on Claude review or GPT for long context, you need to add it.
If none of those apply, the default DeepSeek setup is fine. Do not add complexity for its own sake.
Three Ways to Add a Second Model
| Approach | How it works | Best for | Trade-off |
|---|---|---|---|
| Direct provider key | Open an Anthropic/OpenAI/Google account, paste the key into DSH, pay each supplier directly | Maximum control, lowest per-token price | Multiple keys, multiple invoices, no cross-provider failover |
| Multi-model aggregation endpoint (LumeAPI, OpenRouter, etc.) | One key, one base URL, DSH pulls the model list from the endpoint | Most teams---one dashboard, automatic failover, volume discounts, easy to add a third or fourth model later | Slight per-token markup (typical 20-40% over direct pricing) |
| Local self-hosted model (Ollama, vLLM) | Run a model on your own hardware, point DSH at the local URL | Sensitive data, no API cost, full control | Hardware cost, you operate the model, weaker on hard reasoning tasks |
The rest of this page focuses on the second approach because it is what most teams adopt.
How to Configure Multi-Model Access in DSH
The configuration takes about five minutes. You need DSH running (npx @deepseek-ai/dsh web or a local build) and a key from any OpenAI-compatible multi-model gateway (LumeAPI, OpenRouter, silicon-flow, Volcano, etc.).
- In the DSH web UI, go to Settings -> Models -> Add Provider.
- Select OpenAI Completions as the protocol.
- Fill in the provider form:
- Base URL: Your aggregation endpoint (for example,
https://api.lumeapi.site/v1) - API Key: Your aggregation key
- Name: Any label you prefer
- Click Fetch Available Models. If the connection works, DSH pulls the model list.
- Tick the models you want. Common picks:
gpt-5.5,claude-sonnet-4.6,deepseek-v4-pro,gemini-3.5-flash. - Click Create Provider and return to the home screen.
- In the model selector, you should now see the new models alongside the default DeepSeek models.
- Test with a one-line identity prompt:
What model are you?Each provider should return a different answer.
If the model list fails to load, jump to the troubleshooting section at the end.
Which Model to Pick for Which Task
Once you have several models available, the question becomes: which one for this job? Here is a practical mapping that works for most teams.
| Task | Best model | Why |
|---|---|---|
| Careful code review, security-sensitive changes | Claude Sonnet 4.6 | Lower hallucination, careful reasoning, strong on diff context |
| Large refactor across 100K+ tokens | GPT-5.5 | 1.05M context window, holds whole files in memory |
| PTC mode batch jobs (multi-file migrations) | DeepSeek V4 Pro or GPT-5.5 | Strong step-by-step structured reasoning |
| Quick edits, one-line changes, autocomplete | Gemini 3.5 Flash or DeepSeek V4 Flash | Lowest cost per token, low latency |
| Building a custom DSH preset | Claude Opus 4.7 or GPT-5.5 Pro | Meta-level reasoning about the framework itself |
A common split that works for most teams: 70% Flash-tier for simple work, 25% Sonnet for medium complexity, 5% Opus/Pro for the hardest calls. That mix typically reduces total API spend by 40-60% compared to running everything through a single premium model.
Cost Comparison: Direct vs. Aggregated
The table below shows official per-million-token pricing as of August 2026, side by side with typical aggregated rates available through a multi-model API gateway such as LumeAPI. Official rates are taken from each provider's published pricing page; aggregated rates are typical volume-tier estimates (marked est.) that vary by gateway, tier, and region.
| Model | Official input ($/M tok) | Official output ($/M tok) | Aggregated input ($/M tok, est.) | Aggregated output ($/M tok, est.) | Aggregator discount | Official source (checked 2026-08-19) |
|---|---|---|---|---|---|---|
| Claude Opus 4.7 | $5.00 | $25.00 | $3.50 (est.) | $17.50 (est.) | ~30% | platform.claude.com |
| Claude Sonnet 4.6 | $3.00 | $15.00 | $2.10 (est.) | $10.50 (est.) | ~30% | platform.claude.com |
| GPT-5.5 | $5.00 | $30.00 | $3.50 (est.) | $21.00 (est.) | ~30% | openai.com/api/pricing |
| GPT-5.5 Pro | $30.00 | $180.00 | $21.00 (est.) | $126.00 (est.) | ~30% | openai.com/api/pricing |
| DeepSeek V4 Pro | $2.00 | $8.00 | $1.40 (est.) | $5.60 (est.) | ~30% | platform.deepseek.com |
| Gemini 3.5 Flash | $0.08 | $0.30 | $0.05 (est.) | $0.21 (est.) | ~30% | ai.google.dev/pricing |
How to read the columns - Official input/output: the price you pay when you call the provider's API directly. These are exact numbers from each provider's pricing page. - Aggregated input/output (est.): a typical volume-tier rate you would pay through a multi-model API gateway. The
(est.)tag means the figure is an estimate based on published volume tiers, not a quoted price. - Aggregator discount: the headline percentage you save by routing through an aggregator. This is the price discount only; the bigger savings come from task-based routing (see the 70/25/5 mix above). Always verify live pricing on your chosen gateway before quoting numbers to stakeholders. Aggregator rates depend on volume tier, region, and which models the gateway has negotiated deals for.
Troubleshooting
"Failed to fetch models"
- Verify the base URL ends with
/v1. - Confirm the API key is active.
- Ensure the endpoint supports
GET /v1/models. Some minimal proxies only expose/v1/chat/completions.
Responses are slow
- Geographic distance between your DSH instance and the aggregation endpoint adds latency. Choose an endpoint with a presence in your region.
Model behaves differently than expected
- Some aggregation endpoints rewrite system prompts or route to similar model aliases. Test with an identity prompt (
What model are you?) to verify the actual model in use.
Key is rejected
- Most gateways require a separate key per project. Generate a new key in your gateway dashboard and paste it into DSH.
Frequently Asked Questions
Is DSH just a wrapper around DeepSeek? No. DSH is an Agent execution framework. The model is a plugin. The default happens to be DeepSeek V4 Pro, but you can swap in Claude, GPT, Gemini, or any OpenAI-compatible model.
Does DSH support any OpenAI-compatible endpoint? Yes. Community tutorials have verified this with Volcano Engine, silicon-flow, LumeAPI, and OpenRouter.
Can I use the same aggregation key for DSH and production? You can, but separate keys for development and production make usage tracking and cost alerts easier.
What if one model in the aggregator is down? A well-built aggregator returns an error that DSH surfaces in chat. Some also offer automatic fallback to another model.
Is DSH stable enough for daily use? DSH is a developer preview as of August 2026. It is excellent for experimentation; for production, pin a specific commit and test upgrades in staging.
Sources
- DeepSeek Harness GitHub repository (deepseek-ai/deepseek-harness). Accessed 2026-08-19.
- DeepSeek Harness setup guide, deepseek.day. Published 2026-08-15.
- DeepSeek Harness installation tutorial, Tencent Cloud Developer. Published 2026-08-17.
- DeepSeek Harness model configuration tutorial, wangruofeng007.com. Published 2026-08-17.
- Anthropic API pricing. Accessed 2026-08-19.
- OpenAI API pricing. Accessed 2026-08-19.